Ethmos

ALT Chart from Ethmos Podcast Pulse: bullish vs bearish takes on AI spending (capex) by week, Sep 21 to Oct 25 (weeks of Sep 21, Sep 28, Oct 5, Oct 12, Oct 19). Bullish: 40, 33, 51, 42, 41. Bearish: 29, 19, 29, 41, 52. Episodes that discussed AI spending: 110, 93, 141, 142, 169. Mixed and neutral takes not shown.

ALT 1. Plain English with Derek Thompson (Derek Thompson with Ramp economist Ara Kharazian · 47 min), Oct 23: Who Actually Pays for AI. Ramp sees $275B+ a year of spend at 85,000+ US businesses, $10B+ of it on AI. The median AI buyer spends ~$12–15 per employee per month. The top 10% spend ~$800 per employee per month; the top 1%, $10,000+. 1% of businesses drive ~80% of OpenAI and Anthropic spend on Ramp. Unusually, that 1% is correlated: software, media, finance, technical services. In a market correction, they could all cut spend at the same time.

ALT 2. Plain English with Derek Thompson (Derek Thompson with Ramp economist Ara Kharazian · 47 min), Oct 23: Why Spending Growth at the Top Is Slowing. Top-1% spend growth has slowed over the last couple of months. Not for lack of value: their AI usage keeps rising every month. They route cheap models to simple tasks, top models only to the hardest. Labs compete with sticker-price cuts and long-term contract discounts. Unlike typical software, buyers don’t lock in, which forces more price cuts. Thompson: buyers have become “discount shoppers” among models.

ALT 3. Plain English with Derek Thompson (Derek Thompson with Ramp economist Ara Kharazian · 47 min), Oct 23: Price Wars, Leverage and Who Wins. Claude Code became a massive revenue driver for the American labs. New margin pressure: SpaceX AI, NVIDIA’s own models, overseas rivals. Anthropic’s Fable, then the best model, peaked at ~5–10% of business spend. A slightly cheaper OpenAI model won, with little loss in performance. Businesses now hold the negotiating leverage over model makers. Lab profits get squeezed; compute demand and end users win.

ALT 4. Plain English with Derek Thompson (Derek Thompson with Ramp economist Ara Kharazian · 47 min), Oct 23: Does AI Pay Off, and Can Demand Double?. Firms that adopt AI grow headcount ~10% over the next two years. The gains came entirely from high-intensity users ($50+ per employee per month). Coding agents help the ~3% of Americans who are software engineers. Top US jobs (healthcare, cashiers, food service) are barely augmented. Thompson’s math: model spend must go from ~$200B to ~$400B, then ~$800B. Thompson: “If you stop the clock right now, AI is not a bubble”.

ALT 1. 20VC (Harry Stebbings with Microsoft CEO Satya Nadella · 26 min), Oct 26: Assume the Model Is Compromised. Once thought alignment would “just fall out” of scaling; no longer does. Run models in zero-trust systems that contain a rogue agent before harm. Alignment has to come from training, even pre-training; you can’t “RL away” to it. Agent 365 gives every agent an identity and logs every tool and model call. He wants chain of thought kept in English, with tamper-proof logs. Chain of thought can be unfaithful, so trusted partners should be able to audit it.

ALT 2. 20VC (Harry Stebbings with Microsoft CEO Satya Nadella · 26 min), Oct 26: Open, Closed and “Licensed Weights”. Microsoft’s MAI models won’t be open-weight; think “licensed weights”. The aim: “a simple, straightforward software business,” not network effects. Most token usage will be commodity: open weights or cheap closed models. There will always be a frontier market; science may be the next frontier. Expects 5–6 closed and 5–6 open-weight models, with routers picking per task. Sees more competition in every layer, not one vertically integrated winner.

ALT 3. 20VC (Harry Stebbings with Microsoft CEO Satya Nadella · 26 min), Oct 26: Why Every Firm Needs Its Own Intelligence. Knowledge spread across firms is why capitalism and price discovery work. Arrow’s information paradox flips: now the AI buyer needs protection. Even thumbs-up, thumbs-down signals create value for the model maker. More of AI’s surplus should stay at the firm, in weights, context or private evals. Unilever moved trade claims from a frontier model to an MAI model it controls. In five years, he says, “a firm would be an AI benchmark”.

ALT 4. 20VC (Harry Stebbings with Microsoft CEO Satya Nadella · 26 min), Oct 26: Can AI Lift a 2% World?. Training runs are “just a percentage of R&D,” which runs 15–20% at heavy spenders. Power and silicon shortages are long-lead; “no one knows” when supply catches up. World growth has sat near 2% a year for ages; he thinks AI can inflect it. Copilot in the NHS: better healthcare adds directly to UK GDP. Software demand proved elastic; science and engineering may be next. His dream: AI spreads faster than anything before and “everyone gets lifted”.

ALT Irrational Analysis (Market Memo · Calm Before the IPO Storm), Sep 20. Card 1 of 4: Anthropic’s IPO, Delayed Twice. The expected path: a public S-1 after Labor Day and a late-September listing. It slipped to October, then, per the WSJ, to after the November elections. The author calls that a bad sign and dismisses the safety talk as theater. Why IPO now? Seven finance contacts say private markets can no longer support Anthropic. The author warns of a “desperation IPO” amid the Iran war, the elections and rising rates. The author’s response: less leverage, 64% margin equity vs. a usual 50–55%.

ALT Irrational Analysis (Market Memo · Calm Before the IPO Storm), Sep 20. Card 2 of 4: GPT 6 Astra vs. Fable 5.1 in Coding. The author has seen engineers agree unanimously only twice; usually 5–20% dissent. First, panning GlobalFoundries’ 45CLO node; now, picking Astra over Fable 5.1 for coding. The author’s estimate: Fable 5 held ~80% of coding before GPT 5.6 Sol took some share. Astra then went to near-100% of coding almost instantly at launch. Nearly every programmer the author knows switched, fast and enthusiastically. The author isn’t a programmer and calls this a vibes-based survey.

ALT Irrational Analysis (Market Memo · Calm Before the IPO Storm), Sep 20. Card 3 of 4: Can Coding Alone Justify the AI Trade? The author rejects Altimeter’s All-In slide: AI needing 4% of $30T in knowledge work. Alternative math: triple the ~$2T coding TAM to $6T, give AI 20%, reach the same $1.2T. The author’s vibes-based view: AI is exceptional at coding, and coding alone can carry the trade. AI’s math breakthroughs run through Lean, a language that writes math as formal code. So, the author argues, AI is good at code, and Lean turns math into code. The McKinsey/Altimeter $2T coding TAM is far too low, the author says.

ALT Irrational Analysis (Market Memo · Calm Before the IPO Storm), Sep 20. Card 4 of 4: Margins, ARR and Model Economics. The author says profitability claims at 80–90% gross margins exclude training capex and R&D. The analogy: Nvidia or Intel backing future R&D and new fabs out of earnings. When standard ARR math (28 days × 13) implied slowing growth, bulls switched to July 31 × 365. Rumors the author cites: Anthropic’s models are far less sparse than OpenAI’s. If so, Anthropic is at a disadvantage on both training compute and inference costs. A friend’s guesses: Astra ~6T parameters (300–600B active), Fable ~3T (400B active).

ALT CNBC’s Morning Call (Arjun Kharpal with NVIDIA CEO Jensen Huang · GTC Berlin), Oct 22: Jensen Huang on Neocloud Financing, the Frontier Labs and Memory. Neocloud demand is “off the charts”; what’s missing is working capital. Working to make NVIDIA compute a financeable asset class for lenders. Calls OpenAI and Anthropic the two fastest-growing companies in history. Talk of an Anthropic slowdown is nitpicking that “doesn’t change the fundamentals”. Substrates are the toughest bottleneck; capacity to grow 2.5x in three years. CoWoS packaging capacity to rise “tremendously” through work with Amkor. Doubts a 2029 memory glut: “you could always use more of it”. First to put phone-grade LPDDR memory in data centers, for 2x energy efficiency.

ALT The a16z Show (Ben Horowitz and Erik Torenberg with Amazon CEO Andy Jassy · 61 min), Oct 23: Andy Jassy on Why Amazon Is Investing $220 Billion in AI. AI in AWS: $25B+ run rate in three years; early AWS was ~$60M. ~$220B of 2026 capex still leaves demand “on the table”. AWS capacity: 2026 sold out, 2027 largely, “a good chunk” of 2028. Expects the data center, power and memory crunch to outlast 2027. Inference is “still too expensive” for pervasive use. Graviton: ~40% better price-performance than x86; used by 99% of top 1,000 EC2 clients. Trainium: ~30–40% better price-performance than comparable chips. Predicts companies will rewrite almost all code so agents can update it.

ALT Decoder with Nilay Patel (Nilay Patel with Bill Gates · 77 min), Oct 22: Bill Gates says to prepare for 20 years of ‘tumultuous’ AI disruption. Calls AI “by far the most dangerous technology ever invented”. Expects significant harm from cyberattacks and to jobs within 2–3 years. His big concern: a ~20-year transition with “winners and losers” before abundance. Superhuman AI ends the scarcity of white-collar skills like telesales and accounting. Floats reserving ~40% of jobs for humans, via early retirement and shorter weeks. Three alignment paths: replace RL training, see inside models, or a rule-based “police layer”. A 1 GW AI data center is roughly a $40B investment. Gates Foundation aims to cut child deaths from 5M a year to under 2.5M.

ALT Goldman Sachs Exchanges (Allison Nathan with Goldman’s Daan Struyven · 14 min), Oct 23: Why Oil Prices Are Surging Despite Recovering Supply. Brent ~$112 vs. Goldman’s $85 fair value; recent gains are risk premium. Base case: $90 Brent at year-end, an $80 average in 2027. Risks to that forecast are “clearly skewed to the upside”. Gulf crude exports back to 110% of 2025 levels; diesel only 40%. Crude freight up $20/bbl in a year, ~3/4 of it from rerouting Mideast supply. European gas: €70 if Qatar LNG recovers, €150 with low flows and a cold winter. Copper up 35% YoY, with another 10% upside to $15,600 by end-2027. Central banks are buying gold at 5x the pre-2022 pace.

ALT No Priors (Sarah Guo and Elad Gil with CrowdStrike CEO George Kurtz · 45 min), Oct 22: The New Apex Predator: Adversaries Armed With AI Agents. Average attacker breakout time: 29 minutes; the fastest seen, 27 seconds. Ablated models give hacktivists what nation states had a few years ago. AI hasn’t created a new class of attack; it’s “more and faster” CrowdStrike’s model: ~3 cents per remediation vs. ~$10 for a top frontier cyber model. Its Red Tempest and Blue Solano models were built with NVIDIA, partly on Nemotron. From one module at founding to 33 today. Industry stat: ~90 AI agents per employee, which Kurtz thinks is low. Within two years, local models will run on people’s computers.

ALT Business Breakdowns (With Fraser Christie, founder and CIO of Sandstone Global · 79 min), Oct 22: Affirm: The Internet’s American Express. $50B+ of GMV, growing 30%+ a year, with ~50% of US buy now, pay later. Stock: listed ~$49, peaked ~$175, bottomed at $9, now ~$60–80. ~$25B market cap at ~20x adjusted earnings, excluding stock comp. A 100 bp rate rise costs ~$90M against ~$2.5B of gross profit. ~95% of each quarter’s transactions come from existing customers. The Affirm Card is ~20% of GMV and growing 100%+ a year. Only 75 of the top 250 US merchants use Affirm; the target is $100B GMV. Max Levchin’s salary is under $40,000, with no cash bonus.

ALT 20VC (Harry Stebbings with Rory O’Driscoll and Jason Lemkin · 61 min), Oct 22: Trillion or Bust: The New VC Game. Big-fund venture: “It’s billion-dollar positions or nothing” “CoreWeave is B-tier compared to NVIDIA… Oracle is B-tier” Anthropic Q3: $16.8B of revenue, up 45% on Q2, 85% gross margin. IPO pricing talk slips from $2T to $1.5T. Replit flipped its default from Opus and Fable to OpenAI’s Astra at a 30–40% better price. Modal raised $750M at $15.75B, led by Accel. Anduril at $90B: is it worth 40x revenue? Harvard’s ~$68B endowment got a lift from a ~$2.2B SpaceX stake.

ALT Hidden Forces (Demetri Kofinas and Grant Williams with BCG’s Philipp Carlsson-Szlezak · 78 min), Oct 22: Why Younger Generations Are Convinced They Are Falling Behind. His paper: Gen Z’s median income at 28 is 25% above millennials’ and 50% above boomers’ at that age. On the K-shape: “The rich are getting richer, the poor are not getting poorer” Indexed to 2020, the bottom of the distribution has stronger incomes and balance sheets. Millennial homeownership catches up with Gen X by age 42: delayed, not denied. Real consumption is up $1T+ since COVID while sentiment surveys sit at record lows. Gen Z says success takes $600,000 a year, what 1 in 100 Americans earn. Older generations in the same survey answer closer to $150,000. Gen Z are the “guinea pigs of the smartphone revolution”